胡若轩 孙伟 吴泽毓

摘要：面對传统汽车算法、碰撞测试受限于物理环境及运营资金，无法满足相关领域的训练和验证等问题，本文基于虚拟现实的无人驾驶训练测试仿真平台LEAVS进行优化。LEAVS是一个包含传感器、数据采集以及算法验证模块的自动驾驶平台，基于其优化后的LEAVS二代，核心模块以自动驾驶中包含的人机交互作为切入点，收集真实场景物理数据，构建与真实环境一致测试场景，架设新型传感器反馈体系，完成数据收集与整理框架平台，测试、分析并验证自动驾驶算法，实现一个集自动驾驶汽车测试、问题诊断、险情复现、算法调整于一体的仿真系统。并且支持测试并验证自动驾驶的感知、决策和控制三大主要技术。关键词：自动驾驶；仿真平台Abstract：Faced with the traditional car algorithm，collision test is limited by the physical environment and working capital，can not meet the training and verification issues in related fields，this paper is based on virtual reality unmanned training test simulation platform LEAVS for optimization.LEAVS is an autopilot platform that includes sensors，data acquisition and algorithm verification modules.Based on its optimized LEAVS II generation，the core module uses human-computer interaction included in autonomous driving as an entry point to collect real-world physical data，build and reality.The environment is consistent with the test scenario，a new sensor feedback system is set up，the data collection and finishing framework platform is completed，the automatic driving algorithm is tested，analyzed and verified，and a simulation system integrating autopilot vehicle testing，problem diagnosis，dangerous recurrence and algorithm adjustment is realized..It also supports testing and validating the three main technologies of autopilot perception，decision making and control.Keywords：Autopilot，simulation platform 背景人机交互在桥接网络物理系统层面发挥着重要作用。无人驾驶车辆是这种交互的完美示例，它提供了从环境感知到网络空间智能决策的反馈回路，并实现了无故障决策。在现实世界中进行大量此类测试的代价高昂，通常使用仿真环境进行模拟测试。当前的大部分的仿真环境都支持人工输入数据，但并非所有仿真环境都可以对人类决策进行推理和理解，并推动与不断变化的环境的交互。要进行真正的“自主”驾驶车辆测试，仿真平台应具备自主学习能力。虽然在仿真平台上很容易获得大量数据，但很难处理这些数据，并将这些数据转化为车辆可理解的有效命令，如自动道路引导，环境感知，交通碰撞检测和恢复。……